Technology, Skill and the Wage Structure

Nancy L. Stokey*
University of Chicago and NBER

June 14, 2017 (Revised)

Abstract
Technical change, even if it is limited in scope, can have employment, output, price and wage
effects that ripple through the whole economy. This paper uses a flexible and tractable
framework, with heterogeneous workers and technologies, and many tasks to analyze the general
equilibrium effects of technical change for a limited set of tasks. Output increases and price falls
for tasks that are directly affected. The effects on employment depend on the elasticity of
substitution across tasks. For high elasticities, employment expands to a group of more skilled
workers. Hence for tasks farther up the technology ladder, employment falls, output declines,
and prices and wages rise. For low elasticities, employment at affected tasks contracts among
less skilled workers, as they shift to complementary tasks with unchanged technologies. In all
cases, the output, price and wage changes are damped for more distant tasks, both above and
below the affected group.
JEL Classification: D50, E24, O33, O40
KEYWORDS: technical change, human capital, wages

Department of Economics
University of Chicago
1126 E. 59th Street
Chicago, IL 60637
nstokey@uchicago.edu

*I thank Daron Acemoglu, Ufuk Akcigit, Treb Allen, Marios Angeletos, Enghin Atalay, Gadi
Barlevy, Marco Bassetto, Jess Benhabib, Katarina Borovickova, Paco Buera, Ariel Burstein,
Raquel Fernandez, Xavier Gabaix, Larry Jones, Boyan Jovanovic, Joe Kaboski, Greg Kaplan,
Rasmus Lentz, Jeremy Lise, Bob Lucas, Sydney Ludvigson, Jesse Perla, Steve Redding, Jose
Scheinkman, Ananth Seshadri, Jaume Ventura, Gianluca Violante, and seminar participants at
UCL, the University of Wisconsin, the Federal Reserve Bank of Philadelphia, the Becker-
Friedman Institute, the Stern School, the Federal Reserve Bank of Chicago, and CREi, for useful
comments.
 1. INTRODUCTION

Technical change, even if it is limited in scope, can have employment, output, price
and wage e ects that ripple through the whole economy. This paper uses a exible
and tractable framework, with heterogeneous workers and technologies, and many
tasks/goods, to analyze in detail the general equilibrium e ects of technical change
for a limited set of tasks. Technology and human capital are assumed to be comple-
ments in production, so the labor market|which is competitive|produces positively
assortative matching between technologies and skills: tasks/goods with better tech-
nologies are produced by workers with more human capital. But the quantitative
allocation of workers to technologies is endogenous, determined by demands for the
tasks that are produced. Hence technical change for a limited set of tasks produces
changes in employment, output levels, prices and wages, for tasks and workers not
directly a ected.
Why is a model of this type useful? Not only do wage di erentials across skill or
occupational categories change over time, but even the trends shift. As documented
by Goldin and Katz (2007), and described both succinctly and accurately in their title,
the wage structure in the U.S. has seen \narrowing, widening, and polarizing" over
the last century.1 Explanations for these trends always involve shifts in the relative
supply of and demand for skill, with shifts in demand arising from technological
change. What has been missing from the discussion is a uni ed way to analyze how
technology and skill are matched that is exible enough to accommodate all of these
1
Ober (1948) documents in detail the narrowing of wage di erentials over the period 1907-47;
Goldin and Margo (1992) describe the rapid compression in the 1940's, followed by a slow widening
in the 1950's and 1960's that accelerated in the 1970's and early 1980's; Katz and Murphy (1992)
show that the college premium rose slowly in the 1960's, fell during the 1970's and rose sharply in
the 1980's; and Autor, Katz and Kearney (2006), Machin and Van Reenen (2007), and Autor and
Dorn (2013) document labor market polarization after 1980.

1
trends as possible outcomes. The usual model, which features two skill groups and
labor-augmenting technical change for each group, can explain narrowing or widening
of wage di erentials (with skill-biased or unskilled-biased technical change), but is
inadequate to talk about polarizing.
The model here lls this gap. It has many intermediate goods/tasks, which are
combined to produce a single nal good. Tasks di er in terms of their technology level,
so there is a one-dimensional technology ladder, and workers di er in their human
capital, so there is also a one-dimensional skill ladder. All production functions display
constant returns to scale, and all markets are perfectly competitive, so rms, as such,
play no role. A competitive equilibrium consists of an allocation of skill types to
tasks, and a supporting set of prices and wage rates. Complementarity between skill
and technology implies that the equilibrium features positively assortative matching
(PAM), as in Becker's (1973) classic model of partnership formation.
After an improvement in one technology, a ecting a limited set of tasks, labor
is reallocated across all tasks, and all prices and wage rates change. In the model
here, those e ects can be sharply characterized analytically and easily computed
numerically.
The results are intuitively appealing. First, and unsurprisingly, output increases
and price falls for tasks that are directly a ected by the technical change. General
equilibrium e ects are never strong enough to o set the direct e ect of the shock.
The e ects on employment depend on the elasticity of substitution across tasks and
on the change in relative match quality. To assess match quality, note that because
the equilibrium features PAM, the set of skill levels employed at any particular task
form an interval. Call this the skill bin for that task.
For elasticities across tasks that exceed unity, the substitution e ect works toward
pulling labor into the production of tasks that are directly a ected. Since the change
is an improvement in technology, this e ect is reinforced at the upper threshold of the

2
a ected skill bin. Hence employment necessarily expands to a group of more highly
skilled workers. Consequently employment falls at tasks farther up the technology
ladder, so outputs decline, and prices and wages rise. The e ects are stronger for
tasks with technologies closer to the one enjoying the technical change, and damped
for tasks farther up the ladder.
At the other end of the a ected skill bin, the tendency to pull more labor in is o set
by the fact that the labor is a less suitable match for the newly improved technology.
Either of these forces can dominate, so employment at the lower threshold can expand
to less skilled workers, or it can contract. If it expands, employment falls at tasks
farther down the ladder, so outputs decline, and prices and wages rise. If it contracts,
outputs farther down the ladder increase. In either case the changes are damped for
more distant tasks. The direction of the change at the lower threshold depends in
part on the level of employment at the a ected and neighboring tasks.
For elasticities of substitution across tasks below unity, the previous results are
mirrored and reversed. The substitution force works toward pushing labor away from
tasks that are directly a ected, to increase output of complementary tasks.
At the lower threshold of the a ected skill bin, this e ect is reinforced by the fact
that the less skilled labor at this threshold has become a worse match for the newly
improved technology. Hence employment at the a ected tasks contracts among less
skilled workers. Consequently employment expands for tasks farther down the ladder,
and outputs rise, with damped changes for more distant tasks. Prices and wages may
fall for some tasks and workers closest to those a ected by the technical change.
At the upper threshold of the a ected skill bin, the tendency to push labor out
is o set by the fact that more highly skilled labor is a better match for the newly
improved technology. Hence employment at the a ected tasks can expand or contract.
If it expands, employment falls at tasks farther up the ladder, so outputs decline, and
prices and wages rise. If it contracts, employment and output increase for tasks

3
farther up the ladder. In either case the changes are damped for more distant tasks.
In the cases where the direction of the net e ect is ambiguous, the range of the
a ected skill bin is important in determining the sign. For elasticities that are not
too close to unity and narrow skill ranges, the e ects are rather symmetric up and
down the ladder from the tasks a ected by the technical change. For elasticities that
are close to unity, the signs are ambiguous, but the magnitude of the change is likely
to be small.
As noted above, rms play no role in the analysis, and even the word is (mostly)
avoided. Each worker chooses how to use his labor endowment, combining it with any
of the available technologies. The worker's decision can be viewed as a choice about
an occupation, with his task output used in production of the single nal good.
In some contexts the distinction between human capital and technology is blurred.
Here, human capital is an asset that belongs to a single worker, who is the only one
able to employ it in production. Technology is a nonrival input, used by all workers
producing a particular task. Framed in terms of competitive rms, the technology for
producing a particular task is available to all. However, as will be shown below, the
equilibrium can readily be re-interpreted as one with monopolistically competitive
rms, and the technology as intangible capital that is the property of the producer.
In either case, the fact that it is a nonrival input distinguishes it from both human
and physical capital.
The vast literature on vintage capital models suggests that the distinction between
new technologies and new capital is also blurred. If a new technology requires new
investment for its implementation, giving it one label or the other is largely a matter
of taste. Here, physical capital is ignored, so implementing improved technologies
requires no investment.
The rest of the paper is organized as follows. Section 2 discusses the related liter-
ature. Section 3 presents the basic model and characterizes the competitive equilib-

4
rium. The main results are contained in section 4, where the model is used to study
the e ect of technical change for one set of tasks. In particular, we ask how the labor
allocation, task outputs, task prices and wage rates change, for all tasks and workers.
A su cient condition is provided for the conclusion that \a rising tide lifts all boats,"
that the improvement raises wages for all workers, even those paired with technologies
that are una ected. Section 5 shows how the model can be used to address policy
questions: the e ects of a minimum wage and of opening to international trade or im-
migration. It also analyzes the e ects of eliminating positively assortative matching
and of eliminating technological heterogeneity. Section 6 concludes. Mathematical
arguments and proofs are gathered in the Appendix.

2. RELATED LITERATURE

The model here is related to the extensive theoretical literature on skill-biased

technical change and to the literature on assignment models of the labor market.
The rst models of skill-biased technical change have two types of workers, high-skill
and low-skill, performing distinct and imperfectly substitutable tasks. In particular,
the aggregate supplies of the two types of labor input are inputs into a constant
elasticity of substitution (CES) production function for the single nal good, with
separate (factor-augmenting) technology shocks for each type. Acemoglu (2002) pro-
vides an elegant treatment of this model, and studies its ability to account for some
of the major trends in employment, wages, and skill premia in the U.S.
Acemoglu and Autor (2011), who call it the `canonical' model, provide a nice as-
sessment of its limitations as well as its strengths. They point to four limitations in
particular. First, technical change, whether it is skill-biased or unskilled-biased, nec-
essarily increases the wages of both groups. The model cannot produce wage declines.
Second, because there are only two types of labor, it cannot explain the \polarization"
in the wage structure observed in recent years, as documented in Autor, Katz, and

5
Kearney (2006) and Autor and Dorn (2013). Third, because it does not distinguish
between skills and tasks, it is inadequate for studying the impact of technical change
that a ects only particular tasks. And nally, it cannot explain changes in the allo-
cation of skill groups across tasks. Dealing with the last two limitations requires a
model that distinguishes between skills and tasks.
Acemoglu and Autor (2011) also describe four features that they would like to see
in an alternative to the canonical model. These are: an explicit distinction between
skills and tasks; at least three skill groups; comparative advantage at di erent tasks
across di erent skill groups; and the ability to produce conventional substitution and
complementarity across skill groups. The authors go on to develop a model with
three skill groups (plus capital) and many tasks, with production technologies for
each task that are linear in the four inputs. The factor weights in the linear tech-
nologies are assumed to have the property that higher-skill types have a comparative
advantage in higher-index tasks. A limitation of this setup is that improvements in a
\technology"|a labor-augmentation coe cient in the production functions|a ects
only a single skill-task pair.
Relative to the framework in Acemoglu and Autor (2011), the model here has
many skill groups as well as many tasks. In addition, comparative advantage arises
endogenously as a consequence of the production function, which has skill and task
technology as inputs.
Another strand of the literature on technical change adds physical capital as a factor
of production, as in Autor, Levy, and Murnane (2003) and Autor and Dorn (2013),
and sometimes uses the strong decline in capital (equipment) prices observed in the
data as the technology shock, as in Krusell, et. al. (2000). In these models, physical
capital can enter as a substitute for low-skill or routine labor, reducing its wage rate, or
as a complement to high-skill labor, raising its wage rate. Thus, increasing the supply
of physical capital can produce a wide variety of e ects on wages and employment

6
patterns, depending on the type of capital. The model here has no physical capital.
Although it could be added, the price in terms of tractability is not clear.
The extensive literature on assignment models goes back to Roy (1950), who used
a multidimensional description of ability. In an important contribution, Sattinger
(1975) uses a model very similar to the one here to examine the partial equilibrium
problem of a single employer choosing what types of workers to hire to perform
various tasks. This literature is nicely reviewed in Sattinger (1993). Virtually all of
it is partial equilibrium, while the setup here is a general equilibrium model.
The model here is closest to the one in Costinot and Vogel (2010). Their model,
like the one here, is a general equilibrium setup with one-dimensional heterogeneity
of both workers and tasks. On the technology side, the model here is a special
case of theirs. Speci cally, in the model here skill and technology are inputs into
a CES function with a substitution elasticity (strictly) less than unity, while in CV
the production function is only required to be (strictly) log-supermodular. Thus the
function here satis es the requirement in CV, but the converse does not hold. The
additional assumption brings two important advantages, however.
The rst advantage is that the model here puts no restriction on the type of tech-
nology shocks that can be analyzed. The shocks studied here are limited in scope,
a ecting only one set of tasks. In terms of the distribution function for technologies, a
\simple" shock of this type is a rightward shift over a limited range. Thus, it satis es
rst order stochastic dominance (FOSD).
CV's framework allows them to look at only two types of technology shocks, skill-
biased and extreme-biased. A skill-biased shift requires the relevant distribution
functions to satisfy the monotone likelihood ratio property (MLRP), a stronger con-
dition than FOSD, in general involving shifts throughout the distribution function.
A technology shift in CV is extreme-biased if there exists a threshold technology with
that property that the relevant distribution functions satisfy MLRP above the thresh-

7
old and the reverse condition below the threshold. Again, extreme biased shocks can
be constructed as weighted sums of simple shocks.
Second, the results in CV are only about relative wage e ects, while the model
here delivers conclusions about wage levels, as well output and task price levels. The
relationship between the shocks here and those in CV are discussed in more detail in
section 5.
CV also look at shifts in the distribution of skill that satisfy similar restrictions,
increasing either skill abundance or skill diversity, both de ned using MLRP proper-
ties. The model and methods employed here could be slightly modi ed to study shifts
in the skill distribution. Speci cally, in the setup here the technology distribution is
discrete and the skill distribution is continuous. As will be seen below, this assump-
tion makes it easy to characterize analytically the e ects of a small change in one
technology. A model like the one here, but with discrete skill types and continuous
technologies, could be used to study the e ects of shifts in the supply of skills.
In summary, compared with the literature on biased technical change, the model
here allows extensive heterogeneity in both skills and tasks. Compared with the
assignment literature, the model here is general equilibrium. Compared with Costinot
and Vogel (2010), the CES structure imposed here makes the solution to the general
equilibrium problem easy to characterize, both analytically and numerically, allowing
sharper answers to a wider range of comparative statics questions.

3. THE MODEL

In this section the technologies are described and the competitive equilibrium is
characterized.

8
A. Final good technology

The nal good is produced by competitive rms using intermediate goods/tasks

as inputs. A task is characterized by its technology level xj > 0: There are J such
levels, indexed by j = 1; :::; J; ordered so 0 < x1 < x2 < ::: < xJ : Let j be the share
of tasks with technology level xj : The total number (mass) of tasks is normalized to
unity.
All inputs enter symmetrically into nal good production, but demands for them
di er if their prices di er. In equilibrium, price pj is the same for all tasks with
technology level xj . Hence demand is the same for such tasks. Let yj denote the
(common) quantity for those tasks. The nal good is produced with the constant
returns to scale (CRS) technology
! =( 1)
X
J
( 1)=
yF = j yj ; (1)
j=1

where > 0 is the substitution elasticity. For = 1 the technology is Cobb-Douglas.

The nal goods sector takes the prices pj as given. As usual, input demands are
pj
yj = yF ; all j; (2)
pF
and the price of the nal good is
!1=(1 )
X
J
1
pF = j pj : (3)
j=1

We will take the nal good as numeraire throughout, indexing prices so pF = 1: Input
costs exhaust revenue, so there are no pro ts.
The analysis could be extended to include weights on tasks. Let f!i gIi=1 be a set of
values for the weights, and let ji be the share of tasks with the (technology, weight)
pair (xj ; !i ) : Then output of the nal good is
! =( 1)
X
J X
I
1= ( 1)=
yF = ji !i yeji ;
j=1 i=1

9
where yeji is the input of a task with characteristics (xj ; !i ) : It is straightforward to
show that in this setting prices pj do not depend on i; and demand for each task is

yeji = !i yj ; all i; j;

where fpj ; yj gJj=1 and the aggregates yF ; pF are as above, and

X
I

j ji !i ; all j:
i=1

Output and employment vary with !i across tasks with the same technology xj ; but
the wage structure in the economy depends only on the j 's.

B. Di erentiated good technology

Tasks are produced using heterogeneous labor, characterized by its skill level h; as
the only inputs. Assume that h has a continuous distribution. Let G(h) with density
g(h) on H (hmin ; hmax ) ; with 0 < hmin < hmax 1; denote the distribution of skill
across workers. The total size (mass) of the workforce is normalized to unity, and
labor supply is inelastic: each worker supplies one unit.
The total output of a task depends on the size and quality of the workforce pro-
ducing it, as well as its technology level xj . In particular, if a task with technology xj
employs workers of various human capital levels, with `j (h) 0 denoting the number
(density) of each type, then total output is
Z
yj = `j (h) (h; xj )dh; all j;

where (h; x) is the CES function

=( 1)
(h; x) !h( 1)=
+ (1 !) x( 1)=
; ; ! 2 (0; 1) : (4)

The elasticity of substitution between technology and human capital is assumed to

be less than unity, and ! is the relative weight on human capital.

10
C. Equilibrium

An equilibrium consists of a nal output level yF ; task outputs and prices fyj ; pj gJj=1 ;
a wage function w(h); h 2 H; and an allocation of labor across technologies, that sat-
isfy the usual optimization and market clearing conditions.
The model allows two interpretations about market structure. One is that each
task is produced by competitive rms, with each rm choosing to employ skill types
that minimize unit cost. In this case competition insures that each worker is paid his
marginal revenue product.2 Alternatively, one can suppose that workers simply choose
which task to produce, with each worker choosing a task|a job|that maximizes
his income. In either interpretation, task prices are taken as given by the decision
maker|the rm or the worker.
In principle, the allocation of labor could be quite complicated, with any technology
level xj employing workers with skill h in various disjoint intervals, and with workers
of a given human capital level h producing goods with di erent technologies xj . This
does not occur in equilibrium, and it is straightforward to see why not.
Since labor markets are competitive, the allocation of labor across technologies
is e cient. And since the elasticity of the CES function is less than unity, it is
log supermodular. Hence e ciency requires positively assortative matching: workers
with higher skill h work with higher technologies xj (Costinot, 2009). Consequently
the equilibrium labor allocation is characterized by thresholds hmin = b0 < b1 < ::: <
bJ 1 < bJ = hmax ; where technology xj employs workers with skill h 2 (bj 1 ; bj ) : We
will refer to the interval (bj 1 ; bj ) as skill bin j: An individual with human capital h =
bj is indi erent between working with technologies xj and xj+1 : Since the distribution
2
If > 1; pro t-making rms could be introduced by assuming that each task is supplied by a
unique producer. Under this assumption, the allocation of labor, output of each task, and prices
would be unchanged, but wages would reduced by the factor ( 1) = ; with the residual revenue
going to pro ts.

11
function G is continuous, the set of such workers has measure zero, and they can be
allocated to either bin.
Equilibrium also requires market clearing for goods and labor. Thus, the equilib-
rium conditions are:
a. income maximization by all types of labor,

w(h) pj (h; xj ); all h; with equality if h 2 [bj 1 ; bj ] ; all j; (5)

b. market clearing for tasks: fyj ; pj gJj=1 satisfy (2), with yF as in (1);
c. labor market clearing,
Z bj
(h; xj )g(h)dh = j yj ; all j: (6)
bj 1

The rst condition implies that each task is priced at unit cost, and the last says
that the total productive capacity of labor with skill h 2 (bj 1 ; bj ) is su cient for
production of tasks with technology xj .
The allocation of labor within any skill bin (bj 1 ; bj ) across tasks with technology
xj is, to some extent, indeterminate. Equilibrium determines only the output level
yj , which is the same across tasks with technology level xj . For concreteness we can
suppose that each task is produced by skill types in the interval (bj 1 ; bj ) in proportion
to their representation in the population, but this is not required.3
To characterize the thresholds fbj gJj=11 ; note that (5) implies

w0 (h) h (h; xj )
= ; h 2 (bj 1 ; bj ) ; all j: (7)
w(h) (h; xj )

Hence the equilibrium wage function is piecewise continuously di erentiable, with

J 1
kinks at the points fbj gj=1 :
3
Since has constant returns to scale, the number of rms producing any task|if rm are
introduced into the narrative|is indeterminate. Only the total (productivity-weighted) labor input
and total output are determined in equilibrium.

12
 Since workers with skill bj are indi erent between working with technologies xj and
xj+1 ; it follows immediately from (5) and (2) that

pj+1 (bj ; xj )
= ; (8)
pj (bj ; xj+1 )
yj+1 (bj ; xj+1 )
= ; j = 1; :::; J 1: (9)
yj (bj ; xj )

Unit cost and price are strictly decreasing in j; and output is strictly increasing: goods
with better technologies have lower prices and higher sales. If > 1 (if < 1); total
revenue is increasing in j (decreasing in j):
To characterize equilibrium, combine (6) and (9) to nd that fbj gJj=11 satisfy
Z bj+1 Z bj
j+1 (bj ; xj+1 )
g(h) (h; xj+1 )dh = g(h) (h; xj )dh; (10)
bj j (bj ; xj ) bj 1

j = 1; :::; J 1:

Since < 1; the ratio (bj ; xj+1 )= (bj ; xj ) is strictly increasing in bj : Therefore, since
b0 = hmin is given, for any conjectured b1 ; the sequence fbj gJj=2 de ned recursively by
using (10) is increasing in b1 : Equilibrium requires bJ = hmax : Thus a solution exists
and it is unique.
De ne j to be `total productivity' of labor in the j th skill bin,
Z bj
j (h; xj )g(h)dh; j = 1; :::; J: (11)
bj 1

Then use (6) to write output of each type of good as

1
yj = j; j = 1; :::; J; (12)
j

and write (10) in the more symmetric form

1 1
(bj ; xj+1 ) j+1 = (bj ; xj ) j; j = 1; :::; J 1: (13)
j+1 j

13
D. Skill allocation

To see more clearly how workers and technologies are matched, it is useful to look
at potential wage functions, like those in Neal and Rosen (2000, Figure 3.1). Let
wp (h; xj ) denote the wage a worker with skill h would earn producing a task with
technology xj ;
wp (h; xj ) = pj (h; xj ); all h; all j:

Figure 1 displays potential wages as a function of h; for J = 4 technology levels.

For xed xj ; the potential wage wp (h; xj ) is strictly increasing in h; so each curve
is upward sloping. As a function of xj ; there are two e ects. First, the price pj
is decreasing in xj ; so the intercept decreases with xj : In addition, since is log
supermodular, higher xj implies a steeper slope for as a function of h. Thus, plotted
against h; for various xj values, the potential wage functions cross. A worker's actual
wage is the maximum of his potential wages, as in (5). Hence the wage function w(h)
is de ned by the upper envelop of the four curves, and the crossing points along the
upper envelop are the thresholds bj that divide the skill range into bins.
The four small circles show the choices available to a worker with skill hi : The
potential wage for that worker increases moving from x1 to x2 and from x2 to x3 : But
it falls moving from x3 to x4 , so that worker chooses x3 :
In a model with search frictions, these points would represent rungs on a job ladder
for a worker with skill hi : This worker's rst job might come from an employer of any
type. That rm would pay him his reservation wage, not his marginal revenue prod-
uct, so his initial wage would lie below all of the displayed values. But subsequently,
outside o ers from other rms would raise his wage, for two reasons. If the outside
rm was a better match, he would change jobs and receive a wage increase. But even
if the outside rm was an equivalent (or possibly worse) match, his wage might be
bid up by competition, although in this case he would not change jobs.

14
 4. TECHNICAL CHANGE

This section looks at the e ects of technical change that improves one technology
by a small increment, with all others unchanged. Speci cally, it characterizes the
e ect on the labor allocation, described by the thresholds fbj gJj=11 ; on the output
levels and prices fyj ; pj gJj=1 for all tasks, and on the wage function w(h):
The main forces can be previewed in Figure 1. Suppose technology xk gets the
improvement. The direct e ect is to increase labor productivity for workers in skill bin
k; raising wp ( ; xk ) and making it slightly steeper. But the higher labor productivity
increases yk ; which depresses the price pk , and tends to raise all other prices, pj ; j 6= k:
These price changes lower wp ( ; xk ) partway back toward its original level and raise
all the other curves, wp ( ; xj ); j 6= k: The thresholds de ning the employment bins
shift, changing employment patterns and wages for all workers.
The rest of this section analyzes these changes in detail. Throughout we will use
`hats' to denote proportionate changes induced by the perturbation, z^ z 1 @z=@";
for any variable z: All derivations and proofs are in the Appendix.

A. Final output

Suppose that technical change increases technology xk by a small increment " > 0;
with all others unchanged. Note that the change in output of the nal good yF is a
weighted average of the output changes for tasks,

X
J
y^F = jy
^j ; (14)
j=1

where the weights

1
j j pj yj ; all j; (15)
yF

15
 PJ
with j=1 j = 1; are their cost shares in producing the nal good. With the price
of the nal good xed at unity, the relative price changes for tasks are
1
p^j = (^
yF y^j ) ; all j; (16)
PJ
and the weighted average of the price changes is j=1 jp
^j = p^F = 0:
Consider rst the short run e ects, with labor immobile. Recall the de nition of
j; all j; in (11), and let ^ k be the direct e ect of the technology improvement on
total labor productivity in skill bin k. Output increases for tasks produced with
technology xk ;
1 @ k
y^kSR = ^ k > 0; (17)
k @xk

and is unchanged for all other tasks. Hence the change in nal output is

y^FSR = k
^ k > 0:

In the longer run, with labor mobile, the changes in fyj gJj=1 and yF must be aug-
mented to account for the impact of changes in the skill bins, changes in the bj 's.
n oJ 1
(k)
Let bj (") denote the solution to (13) as a function of ", where b0 = hmin and
j=1
bJ = hmax are xed. De ne the density-weighted changes in the thresholds

(k) (k)0
j g(bj )bj ("); j = 1; :::; J 1; (18)

(k) (k)
with 0 = J = 0: From (11) and (12), the long run changes in output levels for
tasks are
1 h (k) (k)
i
y^k = (bk ; xk ) k (bk 1 ; xk ) k 1 + ^ k; (19)
k
1 h (k) (k)
i
y^j = (bj ; xj ) j (bj 1 ; xj ) j 1 ; all j 6= k:
j

The next proposition shows that, to a rst-order approximation, the change in the
labor allocation has no impact on output of the nal good: the long-run increase is
the same as the short-run increase.

16
 Proposition 1: In the long run, with labor mobile, the change in output of the
nal good is, to a rst-order approximation, the same as in the short run, y^F = y^FSR :

This result is not surprising. The potential additional e ect in the long run arises
only from the reallocation of labor, changes in the thresholds fbj gJj=11 de ning the skill
bins. Since labor markets are competitive, the baseline allocation of labor maximizes
yF : Hence to a rst order approximation, small changes in those thresholds have no
e ect on nal output. An increase (decrease) in bj raises (lowers) the output of tasks
with technology xj ; but the e ect on nal output is exactly o set by the decrease
(increase) in the output of tasks with technology xj+1 :

B. Labor allocation

The changes in the labor allocation do, however, a ect task-level outputs and prices,
as well as wages. The rest of this section describes these changes. To determine the
e ect on the labor allocation, di erentiate (13) and use (11) to get a system of J 1
linear equations for the changes in the thresholds,

(k)
= M A(k) ; (20)

where the superscript denotes which technology has been perturbed, and for any k;

(k) ^x (bk 1 ; xk ) + ^ k ;
Ak 1 = (21)
(k) ^x (bk ; xk ) ^ k;
Ak =
(k)
Aj = 0; otherwise.

Since A(k) has at most two non-zero elements|and only one if k = 1 or k = J; for
(k) (k)
xed k the solution to (20) involves only Ak 1 ; Ak ; and the columns M k 1 and M k :
In particular,

(k) (k) (k)

j = mj;k 1 Ak 1 + mj;k Ak ; j = 1; :::; J 1; (22)

17
where the rst term drops out if k = 1; and the second drops out if k = J:
M is the inverse of a tridiagonal matrix, so it has a recursive structure. Lemma 2
shows that it has strictly positive elements, and that successive row elements above
and below the diagonal have ratios that depend only on the row j:

Lemma 2: All elements of M are positive, and the elements in each column M n
satisfy

mj+1;n = qj+1 mj;n ; j n; (23)

mj 1;n = rj 1 mj;n ; j n;

where fqj+1 gJj=12 and frj 1 gj=2

J 1
are positive constants.

Lemma 2 can be used as follows. Fix k; and use the rst line in (23) to compare
successive rows j +1 > j k in (20). Similarly, use the second line in (23) to compare
successive rows j 1<j k 1; concluding that

(k) (k)
j+1 = qj+1 j ; j k; (24)
(k) (k)
j 1 = rj 1 j ; j k 1:

Thus, all thresholds at and above the k th move in the same direction, and all those
at and below the (k 1)th move in the same direction. It remains to determine the
(k) (k)
signs of k and k 1: For this we need to characterize the two nonzero elements of
A(k) :

Lemma 3: For any k;

(k) (k)
i. if = 1; then Ak 1 > 0 and Ak > 0;
(k) (k)
ii. if > 1; then Ak > 0 and Ak 1 can have either sign; and
(k) (k)
iii. if < 1; then Ak 1 > 0 and Ak can have either sign.

The intuition for Lemma 3 is straightforward from (21). The term ^x (h; xk ) is the
proportionate change in labor productivity for a worker with skill h: Since < 1,

18
it is strictly increasing in h: The term ^ k is the average value of these changes in
skill bin k: If 1; then for a worker with skill bk ; at the upper threshold of the
bin, ^ k < ^x (bk ; xk ) ^x (bk ; xk ); so A(k) > 0: If < 1; the sign is ambiguous.
k

Similarly, if 1; then for a worker with skill bk 1 ; at the lower threshold of skill bin
k; ^x (bk 1 ; xk ) ^x (bk 1 ; xk ) < ^ k ; so A(k) > 0: If > 1; the sign is ambiguous.
k 1

Can anything more be said about the terms with ambiguous signs? The answer
depends, to a large extent, on how the technology levels are chosen/de ned. If the
technology grid is ne, then the skill bins are narrow, so bk 1 is close to bk ; and
(k) (k)
Ak 1 Ak . For = 1; both are close to zero.
(k) (k)
If Ak 1 and Ak are both positive, then it follows immediately from (22) and
Lemma 2 that all thresholds shift upward. But even if one term in (22) is negative,
the sign of the sum can sometimes be determined. Proposition 4 characterize the
(k) (k)
signs of k and k 1 to the extent that it is possible.

Proposition 4: For any k; an increase in technology xk implies:

for = 1;
(k)
j > 0; all j;

for > 1;

(k)
j > 0; j k;

S 0; S 0;
(k) (k)
j j<k 1; as k 1

and for < 1;

(k)
j > 0; j k 1;

S 0; S 0:
(k) (k)
j j > k; as k

For = 1; all thresholds shift upward. For > 1; the thresholds at and above the
k th shift upward, while those at and below the (k 1)th can shift either way. For

19
 < 1; the thresholds at and below the (k 1)th shift upward, while those at and
above the k th can shift either way.

C. Task outputs

From (6) and (11), the change in output for a task of type j 6= k depends on the
sum of the productivity-weighted employment changes at the two thresholds,

(k) 1 h (k) (k)

i
y^j = (bj ; xj ) j (bj 1 ; xj ) j 1 ; j 6= k; (25)
j

(k) (k)
where 0 = J = 0: For goods of type k; the direct e ect of the productivity change
must also be added, so

(k) ^ 1 h (k) (k)

i
y^k = k+ (bk ; xk ) k (bk 1 ; xk ) k 1 : (26)
k

Proposition 5 characterizes the changes in output.

Proposition 5: For any k;

(k)
y^k > 0;

T 0; S 0;
(k) (k)
y^j j > k; as k

T 0; T 0:
(k) (k)
y^j j < k; as k 1

Output rises for tasks of type k: The output change is in the same direction for all
(k)
tasks of type j > k; rising if k < 0; so more labor is devoted to these tasks, and
(k)
falling if k > 0: Similarly, the output change is in the same direction for all tasks
(k) (k)
of type j < k; falling if k 1 < 0 and rising if k 1 > 0: Thus, for 1; output falls
for tasks of types j > k; and for 1; output rises for tasks of types j < k:
Proposition 6 shows that the size of the output changes above and below k are
damped|whatever their sign|for more distant technology types.

20
 Proposition 6: For any k;

(k) (k) (k)

y^1 < y^2 < ::: < y^k 1 ;
(k) (k) (k)
y^k+1 > y^k+2 > ::: > yJ :

D. Prices and wages

Next consider prices and wages. The price of a task rises or falls as its output change
is less than or greater than the output change for the nal good. In particular, from
(16) and Proposition 1,

(k) 1 ^k (k)
p^j = k y^j ; all j: (27)

Proposition 7 describes price changes. For tasks of type k; price falls. For types
j 6= k, price rises if output falls, and the size of the increase is damped for types
more distant from k: The sign of the price change is ambiguous if output rises, but
the price changes are nevertheless ordered, even if there is a sign change somewhere
along the chain. Price decreases, if they occur, are clustered among types near k:

Proposition 7: For any k; an increase in technology xk implies

(k)
p^k < 0:

For j < k;
(k) (k) (k) (k)
0 < p^1 < p^2 < ::: < p^k 1 ; if k 1 < 0;
(k) (k) (k) (k)
p^k 1 < ::: < p^2 < p^1 ; if k 1 > 0;

and some or all of the latter price changes can be negative. For j > k;

(k) (k) (k) (k)

0 < p^J < p^J 1 < ::: < p^k+1 ; if k > 0;

(k) (k) (k) (k)

p^k+1 < ::: < p^J 1 < p^J ; if k < 0;

21
and some or all of the latter price changes can be negative.

Next consider wage changes. It follows immediately from (5) that

(k)
w(h)
^ = p^k + ^x (h; xk ); h 2 (bk 1 ; bk ) ;
(k)
w(h)
^ = p^j ; h 2 (bj 1 ; bj ) ; j 6= k:

For workers in skill bins j 6= k; wages change only because the price of their output
changes. Hence the direction and size of the wage change is the same as the price
change, and is equal for all workers in a skill bin. Workers in skill bin k also experience
a direct productivity e ect, which is increasing in the worker's own human capital h.
Proposition 8 describes the one case where a technology improvement necessarily
raises all wages.
(k)
Proposition 8: If > 1; then for any k; k 1 < 0; implies w(h)
^ > 0; all h:
(k)
If 1; then k 1 > 0; leaving open the possibility that pk 1 falls, so wages fall for
skill bin k 1:
(k)
More generally, if k > 0, then workers in skill bins j > k get wage increases, as
(k)
do workers with human capital near the upper threshold of skill bin k: If k < 0;
wages can fall for some workers at the top of skill bin k: In this case prices can fall
for some or all tasks of type j > k; so that wages fall for workers in these skill bins.
The wage declines are clustered near skill bin k; and are damped for more distant
skill bins. Indeed, wages can rise for workers su ciently far up the skill ladder.
(k)
If k 1 < 0; then workers in skill bins j < k get wage increases, as do workers
(k)
with human capital near the lower threshold of skill bin k: If k 1 > 0; wages can
fall for some workers at the bottom of skill bin k: In this case prices fall for some or
all tasks of type j < k; so that wages fall for workers in these skill bins as well. The
wage declines are clustered near skill bin k; and are damped for more distant skill
bins. Indeed, wages can rise for workers su ciently far down the skill ladder. The
Appendix provides an example where wages decline for some workers.

22
 5. EXAMPLES AND APPLICATIONS

In this section we will rst look at several examples that illustrate the relationship
between the results above and those in Costinot and Vogel (2010). The model will
then be used to study several substantive questions|the e ects of minimum wage
legislation, of opening up to international trade, of positively assortative matching
(PAM), and of technology heterogeneity.

A. Relationship to Costinot and Vogel (2010)

In the model here, the technology space is discrete and the focus is on changes in
a single technology. We will call such shifts simple.
In CV (2010) the technology space is continuous and a shift is described as a change
in the density function weighting various technologies. Speci cally, the technology
values lie in an interval X = [a; b] ; and their weights in the production function for
o
the nal good are represented by a continuous and strictly positive density ( ) on
o n
X: A technology shift changes the density, from to .
CV study two types of shifts. A technology shift is skill biased if the densities
satisfy the monotone likelihood ratio property (MLRP). By de nition, this property
holds if and only if
n n
(x) (x0 )
o (x) o (x0 )
; all x < x0 :

That is, the ratio of the new density to the old must be (weakly) increasing in x:
Lemma 5 in CV shows that after such a shift, every skill type is matched to a (weakly)
better technology. In addition, the two wage functions satisfy MLRP: the propor-
tionate wage increase is larger for higher-skill workers.
A technology shift in CV is extreme biased if there exists a threshold technology
x^ with the property that the densities satisfy MLRP above x^ and satisfy the reverse
property below x^. Lemma 6 in CV shows that after such a shift, there exists a skill

23
threshold h with the property that workers with skill above h are matched with
better technologies and those with skill below h with worse technologies.
To compare the results here with those lemmas, we need to approximate discrete
distributions with continuous densities and vice versa.

i. A continuous approximation to a simple shift.|

J
Fix the discrete technology levels and weights fxj ; j gj=1 ; and consider an incre-
ment of " > 0 to technology k: For the continuous approximation, let a = x1 and
o
b = xJ + (or b = xJ + +"; if k = J), where > 0 is small, and let be a continuous
and strictly positive approximation to f j g. The increment to xk is captured by a
o n
shift in the density to that replaces weight near xk with weight near xk + ":
n
Clearly, such a shift never satis es MLRP. The ratio (x)= o (x) is unity except
near xk ; where it shrinks almost zero, and near xk + "; where it explodes. Hence no
shift of this type satis es CV's de nition of skill biased. The simple shifts considered
in section 4 satisfy rst order stochastic dominance (FOSD), but not MLRP.

ii. Skilled biased shifts.|

As the previous example suggests, in the discrete framework a technology shift
that satis es MLRP requires combining a series of simple shifts. One that is easy
to construct is the discrete approximation to a rightward translation of the density
o o
function. Fix X = [a; b] and : A rightward translation of satis es MLRP if and
only if
D o (x) D o (x0 )
o (x)
> o (x0 )
; all x < x0 ;
o
where D d o =dx: That is D o = o
must be a decreasing function. Suppose this
is the case, so Lemma 5 in CV applies.
o
For the discrete approximation to , choose J large and let " = (b a)=J be the
size of the shift. Let fxj gJj=1 be a uniform grid with step size "; and with x1 = a + "=2

24
and xJ = b "=2: De ne the probabilities f j gJj=1 by
Z xj "=2
o
j = (z)dz; j = 1; :::; J:
xj +"=2

In addition, let xJ+1 = b + "=2 and J+1 = 0:

o
Consider a rightward translation of by ": In the discrete approximation, this
shift changes the probabilities from f j g to f^j g de ned by ^1 = 0; and

^j+1 = j; j = 1; :::; J:

Moreover, this shift clearly is isomorphic to the sum of J simple shifts of the type
described in section 4.
Therefore, summing the changes in (20), the net e ect on the thresholds is
0 1
(1) (2)
A1 + A1
B C
B (2) (3) C
X J
B A + A C
A(k) = M B C;
2 2
=M B .. C
k=1 B . C
@ A
(J 1) (J)
AJ 1 + AJ 1

where
h i h i
(k) (k+1) ^x (bk ; xk ) ^x (bk ; xk+1 ) ^k ^ k+1 ;
Ak + Ak = all k: (28)

As shown in the Appendix, for each k; both of the terms on the right in (28) have
(k) (k+1)
order ": Hence the terms Ak + Ak and the vector also have order ": Recall that
is de ne in (18) as a derivative, so the vector of shifts in the thresholds induced by
a rightward shift of size " in the technology distribution is " : Since itself has order
"; the vector of changes in the thresholds has order "2 :
Does this mean that there is no task upgrading? In the setup here, no change in
the thresholds means that every worker, in every skill bin, works with a technology
that has improved by ": Thus every worker experiences task upgrading. Similarly,
every absolute technology level experiences skill downgrading.

25
 iii. An extreme biased shift.|
For J = 2; a shift that is extreme biased can be constructed from two simple shifts.
Speci cally, an increment of "1 < 0 to x1 together with an increment of "2 > 0 to
x2 satis es the required condition. Let ^b1 denote the new threshold. Workers with
skill below h = ^b1 experience task downgrading, and the complement experience task
upgrading, in accord with Lemma 6 in CV. Workers who remain in the same bin after
the shift experience a change of size j"j j in their technology. If b01 < b1 ; then workers
with skill h 2 [b01 ; b1 ] experience a larger increase. If b1 < b01 ; then workers with skill
h 2 [b1 ; b01 ] experience a larger decrease.

B. Minimum wage legislation

The model here can be used to assess the e ects of minimum wage legislation. The
analysis proceeds by truncating a small segment of workers at the bottom of the skill
distribution, analyzing the e ects on all wages, and then backing out the minimum
wage that would prompt the change. The method is exactly as in section 4, except
that the exogenous shock is a change in the supply of labor of a particular type,
instead of a technology shock.
Fix the baseline economy and its competitive equilibrium, choose " > 0; and sup-
(w)
pose all individuals with h < hmin + " are prohibited from working. Let bj (")
denote the new equilibrium thresholds, as functions of ": The lowest threshold after
(w) (w)
the change is b0 (") = hmin + "; while the top threshold bJ = hmax is unchanged.
The (endogenous) changes in the other thresholds are determined as for a technology
shock.
(w) (w)
Formally, de ne j as the slope of the function bj ("), scaled by the density for
skill at that point,

(w)
0 g(b0 );

26
 (w) (w)0
j g(bj )bj ; j = 1; :::; J 1;
(w)
J 0:

Di erentiate (13) and use (11) to get the analog of (20),

(w)
= M A(w) ;

where M is as before, and here the exogenous shock is the perturbation to labor
supply at the bottom of the skill distribution,

(w) 1 (w)
A1 (b0 ; x1 ) 0 > 0;
1
(w)
Aj 0; j = 2; :::; J 1:

Since A(w) has only one non-zero element, it follows that

(w) (w)
j = mj;1 A1 ; j = 1; :::; J 1:

(w)
Hence by Lemma 2, all the thresholds shift upward, j > 0; j = 1; :::; J 1:
Using (12) and the argument in the proof of Proposition 5, the e ects on task
outputs are

(w) 1 j+1 (w)

y^j = dj +1 (bj 1 ; xj ) j 1 < 0; j = 1; :::; J 1;
j j
(w) 1 (w)
y^J = (bJ 1 ; xJ ) J 1 < 0;
J

J 1 J
where fdj gj=1 and f j gj=1 are de ned in the Appendix. Hence output of every task
falls. Using the argument in the proof of Proposition 6,
(w)
y^j+1 dj+1
j j+2 + j+1 (bj ; xj+1 )
(w)
= cj < 1; j = 1; :::; J 1;
y^j j+1 dj j+1 + j (bj 1 ; xj )

so the changes are damped farther up the task ladder,

j^
y1 j > j^
y2 j > ::: > j^
yJ 1j > j^
yJ j :

27
Using the argument in the proof of Proposition 7, the proportionate price changes
satisfy
p^1 > p^2 > ::: > p^J 1 > p^J :
PJ
Since 1 jp
^j = 0; prices rise for a set of tasks at the bottom of the ladder, and fall
for the complementary set. That is, there exists 1 < k J such that p^j > 0 for
j < k, and p^j 0 for j k: Moreover, the largest price increase is at the bottom of
the task ladder, for j = 1; with more damped changes for 1 < j < k: For tasks at
and above the k th ; the price decreases are larger, farther up the task ladder.
Wages changes exactly parallel the price changes. Thus, wages rise for workers in
skill bins 1 j < k; where price has gone up, with larger increases farther down
the skill ladder. Wages fall for workers in skill bins k and above, with larger declines
farther up the skill ladder.
Finally, notice that the new equilibrium is also the one that prevails if the baseline
economy adopts a minimum wage of

wmin = w(w) (hmin + ") > w(base) (hmin + ") > w(base) (hmin ):

Thus, minimum wage legislation hurts workers at the very bottom of the ladder,
who lose their jobs, bene ts workers in a range above that group, who gain from
the exclusion of close competitors, and hurts workers at the upper end of the skill
ladder, since relative output levels for complementary tasks fall. The workers near
the bottom of the skill ladder who keep their jobs are the biggest winners.
Figure 2 displays, for a numerical example. The substitution elasticity between
technology and skill is set at = 0:5; and the two inputs are given equal weight,
! = 0:5: Four values are used for the elasticity of substitution across tasks, =
0:5; 1:002; 2; and 6. The probability vector for technology types is a discrete
approximation to a Pareto, with shape parameter F = 1:0 and location xmin = 1:
The skill distribution is a truncated lognormal, with mean and variance of unity. In

28
summary, the parameters are

= 0:5; ! = 0:5; 2 f0:5; 1:002; 2; 6g ;

2
F = 1; xmin = 1; h = 1; h = 1:

In each economy the minimum wage is chosen so that it puts 1% of the labor force
out of work. The wage function in these economies is steep at the bottom end, so
the wage at the rst percentile is in each case about 22% higher than at the bottom.
Not surprisingly, in each case the required minimum wage is about 22% higher than
the lowest wage in the baseline economy. Because skill and wage are so low at the
bottom of the skill distribution, the reduction in nal output is between 0.22% and
0.30%, with lower elasticities producing larger losses. Virtually all of the loss is born
by those who lose their jobs: the total wage bill for those who remain employed
is almost unchanged. But in line with the theory, Figure 2 shows that wages rise
slightly for workers who remain employed in the lowest skill bins, and decline slightly
for workers farther up the skill ladder.

C. International trade, immigration

The e ects of international trade, for two special cases, can be analyzed in a similar
way. We will assume that all tasks are costlessly tradeable, and that both countries
produce the same set of tasks and have the same production function for nal output.
Consider a large country whose workers have skills in the range [hmin + ; hmax ] :
Suppose this country opens up to trade with a small, less-skilled partner, one with
workers who have skills in the range [hmin ; hmin + ] : Before opening, the large coun-
try's economy is like the one in the previous section, with the minimum wage. After
opening, the integrated world economy is like the one in the previous section, without
a minimum wage. Hence the e ects on the thresholds, task outputs, prices and wages
for the large country are exactly the reverse of those resulting from a minimum wage.

29
With trade, all bin thresholds in the large country shift downward, so all workers ex-
perience task upgrading. Output of every task is higher in the integrated economy,
and the proportionate output increases are damped farther up the task ladder. Hence
the proportionate price changes satisfy

p^1 < p^2 < ::: < p^J 1 < p^J :

Since the weighted price changes sum to zero, prices fall for a set of tasks j < k; and
rise for the set j k.
Wages follow the same pattern, falling for workers in skill bins where price has gone
down, and rising for those where price has increased. Trade with a skill-poor partner
hurts workers in the lower part of the skill distribution, with the biggest losses at the
bottom of the ladder. Trade bene ts workers at the upper end of the skill ladder,
since output of complementary tasks increases, with the biggest gains going to those
at the top of the skill ladder.
The usual gains-from-trade argument implies that both countries enjoy increases
in total output of the nal good. Hence trade bene ts all the (very similar) workers
in the small, skill-poor country. In the large country, which has more heterogeneity
across workers, there are losers as well as winners.
Figure 2 illustrates the wage e ects for a large, skill-rich country, one with 99%
of the world population, all of whom are at the top of the world skill distribution,
that opens up to trade with a small, poorly-skilled partner, one with 1% of the world
population, all of whom are at the bottom of the skill distribution. For the skill-rich
country, opening to trade results in slight wage reductions in the lowest skill bins and
slight wage gains at the top of the skill distribution, exactly the reverse of the changes
in Figure 2.
The e ects of trade for a large, less-skilled country that opens to trade with a
small, skill-rich partner can be analyzed analytically in exactly the same way. The

30
e ects of opening trade between two countries of similar size or with overlapping skill
distributions is more complex, but could easily be studied numerically.
Finally, in all cases immigration produces exactly the same e ects for workers in
the host country.

D. Gain from positively assortative matching (PAM)

To analyze the gains from positively assortative matching (PAM), it is useful to

introduce rms, and compare economies where skill is and is not observable to rms.
Suppose that each task is produced by many rms, which hire labor and sell output.
Labor and task markets are competitive, so price equals unit cost for all tasks, all
revenue is paid as wages, and there are no pro ts. Then task outputs, task prices
and the wage function are uniquely determined. The number rms and their sizes
are indeterminate, but also irrelevant.
If skill is observable, the equilibrium is exactly as before, and the economy-wide
average wage is E[w(h)] = yF . If skill is unobservable, rms must hire indiscriminately
and pay all workers the same wage, so in the economy with no PAM (NP) the common
wage of all workers is average output, wN P = yF N P : Hence the social gain from PAM,
the increase in the average wage, is the di erence in nal output.
To quantify the gain, we can use a second-order approximation to the production
function for nal goods and, for the NP economy, approximations to the task output
levels. For the NP economy both calculations are straightforward and require no ad-
ditional assumptions. For the economy with PAM, approximations are more di cult.
Thus, we will restrict attention to economies where technology and skill have similar
distributions, so closed form expressions are available. In addition, we require > 1:
To approximate nal output, x a vector (y1 ; :::; yJ ) of task inputs, and let y =
P P
2
j yj ; y = j (yj y)2 ; and cy y =y denote the mean, variance, and coe cient

31
of variation (CoV). Then nal output is
! =( 1)
X ( 1)=
F (y1 ; :::; yJ ) j yj
j

11 2
y 1 c ; (29)
2 y
where the second line approximates around (y; :::; y): Thus, nal output is the mean
of task output, adjusted for its CoV, where the weight on the adjustment is the
inverse of the substitution elasticity across tasks. Hence the change in nal output
from introducing PAM has two components: the change in average task output y;
and the change in the CoV adjustment, the term in parentheses. As will be shown
below, the rst is necessarily positive, but the second can have either sign. Note that
the approximation is good only if the CoV of task output cy is not too large. At a
minimum, we require c2y =2 < :
2 2
Fix the technology and skill distributions, and let (x; x ; cx ) and h; h ; ch denote
the mean, variance, and CoV for each.
For the NP economy, let q(xj ) denote average labor productivity at a task with
technology xj ;
q(xj ) Eh [ (h; xj )] ; all j:

Task outputs in the NP economy are

yjN P = Z(x1 ; :::; xJ )q (xj ); all j;

where ! 1
X
J
1
Z(x1 ; :::; xJ ) iq (xi ) :
i=1
1
Hence labor per task Zq (xj ) is increasing in j: Average skill is the same across
tasks, and better technologies are exploited (only) by allocating more labor to those
tasks. First-order approximations to q and Z imply
q 0 (x)
yjN P q(x) 1 + (xj x) ; all j; (30)
q(x)

32
so task outputs have mean, variance, and CoV

2 2 xq 0
yN P q(x); yN P ( q0) 2
x; cyN P cx : (31)
q
To analyze the equilibrium with PAM, a tractable family of economies are those
with technology and skill distributions that jointly satisfy the following assumption.

Alignment assumption: Let > 1; de ne

=( 1)
1 !
aH = ( 1) ;
!
let fxj g be a ne grid over its whole range, and let f j g and G together have the
P
property that ji=1 i G(aH xj ); all xj :

Under the Alignment Assumption, skill has approximately the same distribution
as technology, but scaled by aH : Thus, h=x = aH ; and ch = cx : Let c denote their
common CoV. For these economies, the competitive equilibrium with PAM has the
property that both the ratio of average skill to technology, aH ; and labor per task,
unity, are approximately constant across tasks.4 Better technologies are exploited
(only) by allocating labor with proportionately higher skill. Hence task outputs are

yjP xj ; all j;

with mean, variance and CoV

2 2 2
yP x ; yP x; cyP c; (32)

where is evaluated at (aH ; 1): Thus, the CoV for task output is the (common) CoV
of the technology and skill distributions.
Under the Alignment Assumption, the mean and CoV of task output in the NP
economy are
1 BAc2 =2
yN P x 1 Ac2 =2 ; cyN P c; (33)
1 Ac2 =2
4
This solution is exact if x has a continuous distribution or h has a discrete distribution.

33
where

1
A( ; ) 2
( 1) > 0; (34)
1
B( ; ) ( 1) ( 2) ;

and B can have either sign. By de nition, both q and q 0 are positive, which requires

c2 =2 < 1=A; and Bc2 =2 < 1=A: (35)

Recall from (29) that PAM a ects nal output by changing the mean and CoV
of task output. Since is strictly concave, the e ect through the mean is always
positive: from (32) and (33),

yN P x 1 Ac2 =2 < x yP :

Clearly the size of the increase is increasing in c: a higher CoV in skill and technology
increases the gain from PAM. And since A is increasing in and decreasing in ;
better substitutability across tasks increases the mean gain from PAM, while better
substitutability between skill and technology reduces the mean gain.
The e ect of PAM through the CoV of task output can have either sign. From (32)
and (33), it reinforces or mitigates the mean e ect,

11 2 11 2
1 c R 1 c ;
2 yP 2 yN P

or equivalently as cyP Q cyN P ; or as

xq 0 1 BAc2 =2
1Q : (36)
q 1 Ac2 =2

For A > 0; the inequalities in (36) hold as B Q 1; where the approximations in (29)
require ( )
2
2 1 BAc2 =2
c =2 < = max 1; : (37)
1 Ac2 =2

34
There are several cases, depending on . In all cases, good approximations to F and
to q; q 0 require c2 to be small. At a minimum, c2 must satisfy (35) and (37).

Case A: As # 1; (34) implies A ! 0 and B ! 1= ; and (36) implies lim !1 cyN P =

cyP : As the production function for nal output converges to Cobb-Douglas, the CoV
e ect contributes nothing to the gain from PAM.
Case B: If 2 (1; 2) ; then B > 1 and cyN P < cyP : In this case the CoV of task
output is smaller in the NP economy, mitigating the gain from PAM.
Case C: If = 2; then B = 1 and cyN P = cyP : In this case the CoV e ect
contributes nothing to the gain from PAM.
Case D: If > 2; then B < 1 and cyN P > cyP : In this case the CoV of task output
is smaller in the economy with PAM, further increasing the gain from PAM.
2
Case E: As ! 1; (34) implies A ! 0 and BA ! ( 1) = < 0; so (36)
implies lim !1 cyN P > cyP : As task inputs become perfectly substitutable, the CoV
adjustment necessarily increases the gain from PAM.

In summary, the CoV adjustment for task output mitigates or reinforces the gain
from the mean e ect of PAM as < 2 or > 2: It is zero at = 2; and it also
vanishes as # 1:
The elasticity of substitution a ects the magnitude of the CoV adjustment. To
see this, consider the two extremes. Note from (36) that lim !1 B = 1; so as the task
production function converges to Cobb-Douglas, the CoV adjustment contributes
nothing to the gain from PAM. At the other extreme,
8
>
> +1; if < 2;
>
<
lim B = 1; if = 2;
!0 >
>
>
: 1 if > 2:

Except in the special case = 2; as converges to Leontief, the size of the CoV
adjustment diverges, with the direction of the e ect depending on :

35
E. Heterogeneous technologies

To assess the e ect of technology heterogeneity on the wage distribution, we can

compare the baseline economy with one that has the same skill distribution, but
a single technology level. To focus on wage inequality, we will choose the single
technology level so that nal output is the same in both economies. Then the total
wage bill is also the same, and only the distribution of wages across workers changes.
With a single technology level, there are many equilibrium skill allocations, but
task outputs, nal output and wages are uniquely determined. In one equilibrium
skill allocation, each task is produced with a pro rata share of all skill levels. For
convenience, we use that one to calculate the common technology level that keeps
nal output unchanged.
To keep nal output yF unchanged, output of each task in the homogeneous tech-
nology (HT) economy must be yjHT = yHT = yF ; all j: Hence the required technology
level xHT satis es
yHT = Eh [ (h; xHT )] :

Then from (2) and (5), the wage change for a worker with human capital h 2 (bj 1 ; bj ) ;
who is in skill bin j if technologies are heterogeneous, is
1=
wHT (h) (h; xHT ) yHT
= ; h 2 (bj 1 ; bj ) : (38)
w(h) (h; xj ) yj

There are two forces, working in opposite directions.

In the baseline economy workers with lower skill are matched with worse tech-
nologies. Hence a worker who in the baseline economy would be in skill bin j; with
xj < xHT ; becomes individually more productive. This e ect works to raise his wage,
the rst term in (38). The reverse occurs for workers in skill bins with xj > xHT ; so
this e ect works to compress the wage distribution.
But task prices also change. In the baseline economy output yj is increasing in xj

36
and price pj is decreasing. Hence a worker who would be in skill bin j; with output
yj < yHT and price pj > pHT = 1; su ers a cut in his product price, the second term
in (38). Output can rise because the technology improves, because the average skill
of co-workers rises, because employment rises, or any combination. This price e ect,
which is reversed for workers in skill bins with yj > yHT ; works to expand the wage
distribution.
The relative strength of the two forces depends on the distributions for technology
and skill. For an analytical assessment it is convenient to consider economies that
satisfy the Alignment Assumption in the previous subsection. Then in the baseline
economy, all workers in skill bin j have human capital of approximately hj aH x j ;
and produce
yj (hj ; xj ) = xj (aH ; 1) ; all j:

In the HT economy they produce (hj ; xHT ): Hence the wage change is

(hj ; xHT ) 1 yHT

ln w(hj ) = ln ln
(hj ; xj ) yj
(aH ; xHT =xj ) 1 yHT xHT
= ln ln
(aH ; 1) xHT (aH ; 1) xj
x 1 x xx x 2 1 1 2
j + j + 0 j j
2 x 2
1 11 2
= 0 2 j; (39)
2

where the second line uses Euler's theorem, the third uses second-order approxima-
tions for (aH ; xHT =xj ) and ln (xHT =xj ) ; the last substitutes for x= and xx = x ;

and where

xHT
1+ j ; all j;
xj
1 yHT
0 ln :
xHT (aH ; 1)

The coe cient on the quadratic term in (39) is negative, and by construction the

37
average wage is unchanged, so 0 > 0: Thus workers with skill near the mean enjoy
a wage gain, and those su ciently far from the mean experience losses.
Figure 3 displays, for the same numerical examples as in Figure 2, the e ects of
eliminating heterogeneity in technologies. In each of the four economies the common
technology xHT is chosen so that nal output (the total wage bill) is unchanged. As
shown in Figure 3, the net e ect in all four economies is to depress wages at both ends
of the skill distribution, and to raise wages for those in the middle. The loss function is
approximately quadratic, as (39) suggests, even though here the distribution functions
for skill and technology are very di erent from each other.
Interestingly, in every case the variance of log wages in the HT economy is slightly
higher than in the baseline economy. In these examples technology inequality reduces
wage inequality, because of substantial price e ects.

6. CONCLUSION

The analysis here has focussed on the e ects of technology changes, but the frame-
work could also be extended and used to examine other questions. As illustrated
in the examples in section 5, it could be used to study the wage shifts arising from
changes in minimum wage, trade, and immigration policies.
It could also be used to revisit the role of labor market frictions in generating un-
employment and producing job ladders. The model here is close to the one in Lise,
Meghir, and Robin (2016), which also uses a framework with heterogeneous work-
ers and technologies, and a CES production function that combines the two inputs.
Relative to that model, the one here drops search frictions, but endogenizes the task
prices|output prices across worker-technology pairs. Here there is a downward slop-
ing demand curve for each task, and its position depends on nal good production.
This fact produces interactions between the wages of di erent workers employed at
the same task and at di erent tasks. Closing the model in this way provides a micro-

38
foundation for the match surplus function, a function that frictional search models
take as exogenous. As a consequence, the model here produces a non-degenerate dis-
tribution of workers across technologies/tasks, even in the absence of search frictions.
Thus, it o ers a richer framework for asking how important frictions are in generating
wage di erentials across workers.
In the framework here, individuals work in isolation to produce tasks outputs. But
most goods and services, whether for consumption or investment, are not produced
by single individuals. Aggregating tasks into goods requires additional information
about which tasks are involved and how they are combined|a better understanding
of what goes on inside rms. And since a rm may produce only one task or a wide
variety of goods, these questions also require thinking about the boundaries of a rm,
about the choices of which set of tasks/goods/services to sell in the marketplace,
which tasks to produce in-house, and which tasks to purchase in the marketplace.
Tackling these questions is important for connecting the job/occupation decisions of
individual workers with the outputs of goods/services measured in most data sources.
Moreover, the patterns for recent wage changes suggest rather strongly that rms are
5
important in determining how technical change gets translated into rising wages.
Wage inequality has displayed large and long-lived shifts over the last century, as
described in Goldin and Margo (1992), Goldin and Katz (2007, 2008), and Autor and
Dorn (2013), and many of these shifts are surely due to changes in technology. Large
increases in wage inequality lead, understandably, to calls for policies to deal with
it. But to such formulate policies, we rst need to better understand the underlying
sources of wage inequality.

5
For example, see Song, et. al. (2016).

39
 REFERENCES

[1] Acemoglu, Daron. 2002. Technical change, inequality, and the labor market, Journal
of Economic Literature, 40(1): 7-72.

[2] Acemoglu, Daron, and David Autor. 2011. Skills, tasks and technologies: implications
for employment and earnings. In Handbook of Labor Economics, Vol. 4, Part
B, edited by Orley Ashenfelter and David Card, pp. 1043-1171. Amsterdam:
Elsevier.

[3] Autor, David H. and David Dorn. 2013. The growth of low-skill service jobs and
the polarization of the US labor market, American Economic Review, 103(5):
1553-1597.

[4] Autor, David H., Lawrence F. Katz, and Melissa S. Kearney. 2006. The polarization
of the U.S. labor market, American Economic Review: Papers and Proceedings,
96(2): 189-194.

[5] Autor, David H., Frank Levy, and Richard J. Murnane. 2003. The skill content of
recent technological change: an empirical exploration, Quarterly Journal of Eco-
nomics, 118(4): 1279-1333.

[6] Becker, Gary S. 1973. A theory of marriage: part I, Journal of Political Economy,
81(4): 813-846.

[7] Costinot, Arnaud. 2009. An elementary theory of comparative advantage, Economet-

rica 77(4): 1165-92.

[8] Costinot, Arnaud, and Jonathan Vogel. 2010. Matching and inequality in the world
economy, Journal of Political Economy, 118(4): 747-786.

40
 [9] Goldin, Claudia, and Lawrence F. Katz. 2007.Long-run changes in teh wage structure:
narrowing, widening, polarizing, Brookings Papers on Economic Activity 2: 135{
165.

[10] Goldin, Claudia, and Lawrence F. Katz. 2008. The Race between Education and Tech-
nology, Harvard University Press.

[11] Goldin, Claudia and Robert Margo. 1992. The great compression: the wage structure
in the United States at mid-century, Quarterly Journal of Economics, 107(1):
1-34.

[12] Huang, Y. and W. F. McColl. 1997. Analytical inversion of general tridiagonal ma-
trices, Journal of Physics A: Mathematical and General, 30: 7919-33.

[13] Katz, Lawrence F. and Kevin M. Murphy. 1992. Changes in relative wages, 1963-1987:
supply and demand factors, Quarterly Journal of Economics, 107(1): 35-78.

[14] Krusell, Per, Lee E. Ohanian, Jose-Victor Rios-Rull, and Giovanni L. Violante. 2000.
Capital-skill complementarity and inequality: a macroeconomic analysis, Econo-
metrica 68(5): 1029-1053.

[15] Lise, Jeremy, Costas Meghir, and Jean-Marc Robin. 2016. Matching, sorting and
wages, Review of Economic Dynamics, 19: 63-87.

[16] Machin, Stephen, and John Van Reenen. 2007. Changes in wage inequality, Center
for Economic Performance Special Paper No. 18.

[17] Neal, Derek, and Sherwin Rosen. 2000. Theories of the distribution of earnings, Chap-
ter 7. Handbook of Income Distribution, ed. by A.B. Atkinson and F. Bour-
guignon, Elsevier Science.

41
[18] Ober, Harry. 1948. Occupational wage di erentials, 1907-1947, Monthly Labor Review,
67(2): 127-134.

[19] Roy, A.D. 1950. The distribution of earnings and individual output, Economic Jour-
nal, 60: 489-505.

[20] Sattinger, Michael. 1975. Comparative advantage and the distributions of earnings
and abilities, Econometrica 43(3): 455-468.

[21] Sattinger, Michael. 1993. Assignment models and the distribution of earnings, Journal
of Economic Literature, 31(2): 831-880.

[22] Song, Jae, David J. Price, Fatih Guvenen, and Nicholas Bloom, Till von Wachter,
Firming up inequality, working paper, October 2016.

42
 APPENDIX

A. Proof of Proposition 1

Proof of Proposition 1: Use (19) in (14) to nd that

X 1 h i
J
(k) (k)
y^F = k ^ k + j (bj ; xj ) j (bj 1 ; xj ) j 1 :
j=1 j

Hence it su ces to show that

X
J 1
j p j yj j+1 pj+1 yj+1 (k)
0 = (bj ; xj ) (bj ; xj+1 ) j
j=1 j j+1

X
J 1
j yj j+1 yj+1 (k)
= w(bj ) j
1 j=1 j j+1

X
J 1
j (bj ; xj ) j+1 (bj ; xj+1 ) (k)
= (bj ; xj ) yj w(bj ) j ;
1 j=1 j j+1

(k) (k)
where the rst line uses (15) and the fact that 0 = J = 0; the second uses (5),
and the third uses (9). From (13), the term in brackets in the last line is zero, for all
j.

B. Matrix M and proofs of Results 2 - 8

Di erentiate (13) and use (11) to get

(k) 1 (k)
Aj = (bj 1 ; xj ) j 1
j
1 1 (k)
+ (bj ; xj ) + (bj ; xj+1 ) + j j
j j+1
1 (k)
(bj+1 ; xj+1 ) j+1 ; j = 1; :::; J 1:
j+1

Write this in matrix form as

(k)
A(k) = T ;

43
where A(k) is de ned in (21), and T is a tridiagonal matrix of dimension (J 1) ;
with rows (0; :::; 0; cj ; aj ; dj+1 ; 0; :::; 0); where

aj j (dj + cj+1 ) > 0; j = 1; :::; J 1; (40)

h i
j
^h (bj ; xj+1 ) ^h (bj ; xj ) =g(bj ) > 0; j = 1; :::; J 1;
1
cj (bj 1 ; xj ) < 0; j = 2; :::; J;
j
1
dj (bj ; xj ) < 0; j = 1; :::; J 1:
j

1
The matrix in (20) is the inverse, M = T :
To characterize M; de ne the constants f i gJi=01 ; f i gJi=1 ; by

0 1; 1 a1 ; (41)

i ai i 1 ci di i 2; i = 2; :::; J 1;

J 1; J 1 aJ 1; (42)

i ai i+1 ci+1 di+1 i+2 ; i=J 2; :::; 1:

Lemma A1 shows that these constants and certain sums are positive.

Lemma A1: The constants satisfy i > 0; all i; and i > 0; all i; and in addition

i 1 + ci i 2 > 0; i = 2; :::; J 1; (43)

i + di i+1 > 0; i=J 2; :::; 1: (44)

Proof of Lemma A1: Use (40) in (41) to nd that

i + ci+1 i 1 = i i 1 di ( i 1 + ci i 2) ; i = 2; :::; J 1:

Since i > 0; di < 0; ci+1 < 0; all i; it follows that

i 1 > 0 and i 1 + ci i 2 >0 =) i + ci+1 i 1 > 0 and i > 0:

44
Since 1 = a1 > 0; and
1 + c2 0 = a1 + c2 > 0;

by induction (43) holds. Similarly, use (40) in (42) to nd that

i + di i+1 = i i+1 ci+1 ( i+1 + di+1 i+2 ) ; i=J 2; :::; 1;

so

i+1 > 0 and i+1 + di+1 i+2 >0 =) i + di i+1 > 0 and i > 0:

Since J 1 = aJ 1 > 0 and

J 1 + dJ 1 J = aJ 1 + dJ 1 > 0;

by induction (44) holds.

Proof of Lemma 2: The matrix M has elements (see Huang and McColl, 1997)
1
mnn = n 1 n+1 ; n = 1; :::; J 1; (45)
J 1
j+2
mj+1;n = cj+1 mj;n ; j = n; :::; J 2;
j+1
j 2
mj 1;n = dj mj;n ; j = n; :::; 2:
j 1

Since i; i > 0 and di ; ci < 0, all i; clearly mjn > 0; all j; n. In addition, clearly the
columns satisfy (23), where

j+2
qj+1 cj+1 ; j = 1; :::; J 2; (46)
j+1
j 2
rj 1 dj ; j = 2; :::; J:
j 1

Proof of Lemma 3: From the de nitions of ^x and ^ x ;

R bk
(k) (b
x k ; x k ) bk 1 x
(h; xk )g(h)dh
Ak = R bk ;
(bk ; xk ) (h; xk )g(h)dh bk 1

45
and since is a CES function,

1=
x (h; x) = (1 !) x (h; x)1= :

Hence Ak T 0 as
(k)

Z bk
(h; xk ) (bk ; xk )1= 1
(h; xk )1= 1
g(h)dh T 0: (47)
bk 1

T 0 as
(k)
An analogous argument (with careful attention to signs) establishes that Ak 1

Z bk
(h; xk ) (bk 1 ; xk )1= 1
(h; xk )1= 1
g(h)dh S 0: (48)
bk 1

Recall that ( ; xk ) is increasing in its rst argument, and < 1: For 1; the term
(k)
in square brackets in (47) is positive over the range of integration, so Ak > 0: For
(k)
1; the term in square brackets in (48) is negative, so Ak 1 > 0. In other cases
the signs are ambiguous.

Proof of Proposition 4: For = 1; the claims are immediate from (22) and
Lemmas 2 and 3. For 6= 1; the same is true for k = 1 and k = J; since (22) has only
one term.
For 6= 1 and k 6= 1; J; use the rst line of (23), with j = n = k 1; in (22) to nd
that

(k) (k) (k)

k = qk m k 1;k 1 Ak 1 + mk;k Ak
k+1 (k) (k)
= ck k 2 Ak 1 + k 1 Ak ; (49)
J 1

where the second line uses (45) and (46). Similarly, use the second line of (23), with
j = n = k; in (22) to nd that

(k) (k) (k)

k 1 = mk 1;k 1 Ak 1 + rk 1 mk;k Ak
k 2 (k) (k)
= k Ak 1 dk k+1 Ak : (50)
J 1

46
 (k)
Suppose > 1: Then Ak > 0; so the second term in (49) is positive. If in addition
(k) (k) (k)
Ak 1 0; then the rst term is nonnegative, so k > 0: If Ak 1 < 0; then
Z bk Z bk
1= 1= 1
0 < (h; xk ) g(h)dg < (bk 1 ; xk ) (h; xk )g(h)dh
bk 1 bk 1
Z bk
< (bk ; xk )1= 1
(h; xk )g(h)dh;
bk 1

(k) (k)
so Ak 1 < Ak : Hence by Lemma A1 the sum in parenthesis in (49) is positive. In
(k) (k)
(50), the fact that Ak 1 < Ak ; does not help in applying Lemma A1, so the sign
is ambiguous.
(k)
Similarly, suppose < 1: Then Ak 1 > 0; so the rst term in (50) is positive. If in
(k) (k) (k)
addition Ak 0; then the second term is nonnegative, so k 1 > 0: If Ak < 0; then
Z bk Z bk
1= 1= 1
(h; xk ) g(h)dg > (bk ; xk ) (h; xk )g(h)dh
bk 1 bk 1
Z bk
1= 1
> (bk 1 ; xk ) (h; xk )g(h)dh > 0;
bk 1

(k) (k)
so Ak < Ak 1 : Hence by Lemma A1 the sum in parenthesis in (50) is positive. In
(k) (k)
(49), the fact that Ak < Ak 1 ; does not help in applying Lemma A1, so the sign
is ambiguous.

Proof of Proposition 5: Recall from (40) that

(bj ; xj )cj = (bj 1 ; xj )dj ; all j: (51)

For j > k; use the rst line in (24) in (25) to nd that

(k) 1 (k)
y^j = [ (bj ; xj )qj (bj 1 ; xj )] j 1
j
1 j+1 (k)
= cj (bj ; xj ) (bj 1 ; xj ) j 1
j j
1 j+1 (k)
= dj +1 (bj 1 ; xj ) j 1 (52)
j j

T 0 S 0;
(k)
as k

47
where the second line uses the de nition of qj ; the third uses (51), and the last uses
Lemma A1 and Proposition 4. Similarly, for j < k, use the second line in (24) in
(25), and the de nition of rj 1 to nd that

(k) 1 (k)
y^j = [ (bj ; xj ) rj 1 (bj 1 ; xj )] j
j
1 j 2 (k)
= (bj ; xj ) + dj (bj 1 ; xj ) j
j j 1
1 j 2 (k)
= 1 + cj (bj ; xj ) j (53)
j j 1

T 0 T 0;
(k)
as k 1 j < k:

For j = k; the rst term in (26) is clearly positive. If 1; then the second term
(k) (k)
is also positive. If in addition k 1 0; then last term is nonnegative, and y^k > 0.
(k)
If k 1 > 0; use the fact that equilibrium requires

(bk 1 ; xk 1 )pk 1 = (bk 1 ; xk )pk ;

before and after the shock. Hence

h i (k)
p^k 1 p^k = ^h (bk 1 ; xk ) ^h (bk 1 ; xk 1 ) k 1
+ ^x (bk 1 ; xk ): (54)
g(bk 1 )
(k)
For k 1 > 0; both terms on the right are positive, so p^k < p^k 1 : Hence y^k > y^k 1 ;
(k)
and as shown above, in this case y^k 1 > 0:
(k)
If < 1; then the rst and third terms in (26) are positive. If in addition k
(k) (k)
0; then second term is nonnegative, and y^k > 0. If k < 0; use the fact that
equilibrium requires
(bk ; xk )pk = (bk ; xk+1 )pk+1 ;

before and after the shock. Hence

h i (k)
p^k p^k+1 = ^h (bk ; xk+1 ) ^h (bk ; xk ) k ^x (bk ; xk+1 ):
g(bk )

48
 (k)
For k 1 < 0; both terms on the right are negative, so p^k < p^k+1 : Hence y^k > y^k+1 ;
(k)
and as shown above, in this case y^k+1 > 0:
(k) (k)
Proof of Proposition 6: For j > k; use (52), the fact that j = j 1 = qj and
the de nition of qj to nd that
(k)
y^j+1 cj+1 dj+1 j+2 = j+1 +1 j+1
(k)
= cj
y^j cj dj j+1 = j +1 j

cj+1 (dj+1 j+2 + j+1 )

=
dj j+1 + aj j+1 dj+1 cj+1 j+2
cj+1 (dj+1 j+2 + j+1 )
= < 1; j > k;
j j+1 cj+1 ( j+1 + dj+1 j+2 )

where the second line uses the de nitions of j; the third uses the de nition of aj ;
and the inequality follows from Lemma A1 and the fact that cj+1 < 0:
(k) (k)
Similarly, for j < k; use (53), the fact that j 1= j = rj 1 and the de nitions of
rj 1 ; j 1; and aj 1 to nd that
(k)
y^j 1 dj 1 1 + cj 1 j 3 = j 2 j 2
(k)
= dj
y^j dj 1 + cj j 2 = j 1 j 1

dj
( j 2 + cj 1 j 3 )
1
=
aj 1 j 2 dj 1 cj 1 j 3 + cj j 2
dj 1 ( j 2 + cj 1 j 3 )
= < 1; j < k:
j 1 j 2 dj 1 ( j 2 + cj 1 j 3)

Proof of Proposition 7: For j 6= k; the claims are immediate from (27) and
Propositions 5 and 6.
(k) (k) (k)
For j = k there are two cases. If k 1 > 0; then y^k 1 > 0 and p^k 1 < 0: Since both
(k) (k)
terms on the right in (54) are positive, it follows that p^k < p^k 1 < 0: This argument
(k)
always holds if 1; and holds for > 1 if k 1 > 0:
(k) (k) (k)
If > 1 and k 1 < 0; then y^j < 0 and p^j > 0; all j < k: In addition, since
(k) (k) (k) J (k)
k > 0; in this case y^j < 0 and p^j > 0; all j > k: Since j=1 j p
^j = 0; it follows
(k)
that p^k < 0:

49
 Proof of Proposition 8: For h 2
= (bk 1 ; bk ) ; the claim is immediate from
Propositions 4 and 7. For skill bin k; note that w(b
^ k 1 ) = p^k 1 > 0; and w(h)
^ is
increasing in h for h 2 (bk 1 ; bk ) :

C. An example with wage declines

For an example where wages fall for some workers, let J = 3 and k = 2; and let
the skill distribution be discrete, also with three types. Let hi ; `j ; j = 1; 2; 3; be the
skill types and the number of workers of each type. The parameters are

x3 = 10; 000; x2 = 4; x1 = 1; x02 = 1:01x2 ;

h3 = 10; 000; h2 = 4; h1 = 0:95;
3 = 0:99; 2 = 0:0090; 1 = 0:0010;
`3 = 0:988912; `2 = 0:007991 `1 = 0:003097
= 0:22; ! = 0:5; = 1:2:

The vast majority of rms have technology x3 ; and the vast majority of the workforce
has skill h3 = x3 ; and these levels are much higher than the others. Hence the increase
in technology x2 leaves nal output virtually unchanged, and the price change at x1
rms depends almost entirely on their own output change. In the initial equilibrium
all workers with skill h3 are employed at rms with technology x3 ; and all with skill
h2 are matched with technology x2 : Workers with skill h1 are divided between rms
with technologies x1 and x2 : The increase in x2 reallocates some additional h1 workers
to x1 rms, and p1 falls. Workers with skill h1 take a wage cut equal to decline in p1 .

D. Skill-biased technical change

Here we will show that for each k; both of the terms on the right in (28) have order
": For the rst term, note that by construction

^x (bk ; xk ) ^x (bk ; xk+1 ) = ^x (bk ; xk ) ^x (bk ; xk + ")

50
 x x +" xx
+" x
x x xx
" ;
x

where and its derivatives are evaluated at (bk ; xk ) :

For the second term, rst note that
R bk R bk+1
x (h; xk )g(h)dh x (h; xk + ")g(h)dh
^ k ^ k+1 = Rbk 1 b
R kbk+1
bk
bk 1
(h; xk )g(h)dh bk
(h; xk + ")g(h)dh
x (hk ; xk ) x (hk+1 ; xk ) + " xx (hk+1 ; xk )
;
(hk ; xk ) (hk+1 ; xk ) + " x (hk+1 ; xk )
where Z bk
hk hg(h)dh; all k;
bk 1

is the average value in skill bin k: To approximate hk+1 in terms of hk ; let H(x) denote
the inverse matching function in the continuous framework: technology x is paired
with skill H(x): Then by construction, hk H(xk ); all k; so

hk+1 hk + "H 0 (xk ); all k:

Hence

^k ^ k+1 x + " [ xh H 0 + xx ]
x
+ " ( hH 0 + x)
0
x ( hH + x) ( xh H 0 + xx )
"
[ + " ( h H 0 + x )]
0 0
x hH + x xh H + xx
" ;
x

so this term also has order ":

E. The gain from PAM

For the approximation in (29), note that F (y; :::; y) = y; and

!1=( 1)
X ( 1)= 1=
Fj = k yk j yj jy = j ; all j;
k

51
 1 1= 1= 1
Fji = y (2 )=
j yj i yi jy = y 1
j i; all i 6= j;
1 1 2 1 1 1=
Fjj = y j y 1= j yj jy
1 1 2 1 1
= y j y j; all j:

Hence
X
Fj (yj y) = 0;
j
XX 1 2
y
Fji (yj y) (yi y) = :
j i
y

For the approximation in (31), note that Z (x; :::; x) = q (x)1 ; and Zi = z0 i ; all
P
i; where z0 > 0 is a constant. Hence i Zi (xi x) = 0; and yjN P is as in (30).
For the approximations to q and q 0 ; use Euler's theorem to nd that
" 2
#
2
h c
hh h
q(x) x 1+ ;
x 2
" 2
#
2
h hhx c h
q 0 (x) x 1+ ;
x x 2

where and its derivatives are evaluated at h=x; 1 : Under the Alignment Assump-
tion h=x = aH : Then by straightforward calculation

a2H hh = = A; a2H hhx = x = BA;

where A; B are as in (34), and x= = 1. Hence xq 0 =q is as in (36).

F. Wage e ects of heterogeneous technologies

For economies that satisfy the Alignment Assumption, the wage change from elim-
inating heterogeneity in technologies is
1
ln w(hj ) 0 ln xH ln xj + ln aH ; xH =xj ln (aH ; 1) ;
1 1 2
0 xj + "x xj + "xx xj ;
2

52
where
@ ln (h; x) @ x x(h; x)
"x ; "xx :
@ ln x @ ln x (h; x)
For the elasticities, note that

=( 1)
(h; x) !h( 1)=
+ (1 !) x( 1)=

x (h; x) = (1 !) x 1= (h; x)1= ;

1 x
xx (h; x) = (1 !) x 1= (h; x)1= x 1
;

so

x x ( 1)=
= (1 !) (h=x; 1) ;

x2 xx 1 ( 1)= x x
= (1 !) (h=x; 1) 1 :

Note, too, that

( 1)= 1
(aH ; 1) = :
(1 !)
Hence evaluating the elasticities at (h; x) = (aH ; 1) gives

x x ( 1)= 1
"x = = (1 !) (aH ; 1) =

" #
2
x x x xx
"xx = x x +

1 1 1 ( 1)= 1
= 2
+ (1 !) (aH ; 1) 1
1 1
= 2
:

53
 Figure 1: potential wages

1.4

1.3 x1 < x2 < x3 < x4 wp(h,x4)

1.2
o
o
1.1 wp(h,x3)
o
ln(wp)

o
0.9

0.8 wp(h,x2)

0.7

0.6 wp(h,x1)

0.5
b1 b2 hi b3
log(h)
 -3
x 10 Figure 2: wage changes from minimum wage legislation
3.5

3 =6

=2
2.5
 = 1.002
ln(wMW) - ln(w)  = 0.5
2

1.5

0.5

-0.5

-1 1% of workforce unemployed

-1.5
-1 -0.5 0 0.5 1 1.5 2 2.5 3
ln(h)
 Figure 3: wage changes from homogeneous technologies

0.05

-0.05

-0.1 ln(wHT) - ln(w)

-0.15

-0.2 =6

=2
-0.25
 = 1.002

 = 0.5
-0.3

-1 -0.5 0 0.5 1 1.5 2 2.5 3

ln(h)